Skeleton detection inference failed! - CUDA Error when using Fusion - Neural Mode

pachermann · November 23, 2025, 6:43pm

we have now enhanced the Fusion Setup and are able to wait for Cans and start the fusion more safely.

However now when we have to fusion running we encounter after a while a total crash of the fusion.

We tried then to use the unity experience and the fusion works very good, stable and synch below 1.0, until the problem encounters.

Its unclear if we just imagine it but as soon as we start the Experience where unity is getting the fused data the crash of the fusion starts appearing after the experience runs about 4-5 min.

We are not sure if this is related to the unity Experience that is running, but it kinda looks like ven do we thought unity is only a reciever of the fuded data.

We tried using both setup types of the new SDK

12 & 13 and we have the same Crash on both versions of the SDK on Microsoft Windows 11 Pro, Version 10.0.26100.

The error keeps repeating when it happens:

[2025-11-23 18:47:37 UTC][ZED][ERROR] Skeleton detection inference failed![2025-11-23 18:47:37 UTC][ZED][ERROR] CUDA ERROR in enum sl::ERROR_CODE __cdecl sl::grab(struct sl::RuntimeParameters)
in enum sl::ERROR_CODE __cdecl sl::Mat::updateGPUfromCPU(struct CUstream_st *,int) : Err [700]: an illegal memory access was encountered.
in void __cdecl ai::BatchedBuffer::memcpy_d2h(unsigned __int64,struct CUstream_st *) : Err [700]: an illegal memory access was encountered.

in void __cdecl ai::BatchedBuffer::memcpy_d2h(unsigned __int64,struct CUstream_st *) : Err [700]: an illegal memory access was encountered.
CUDA error at C:\builds\sl\ZEDKit\lib\src\sl_zed\DetectionHandler.cpp:2261 code=700(cudaErrorIllegalAddress) “cudaEventRecord(ev_bt, cu_strm)”
CUDA error at C:\builds\sl\ZEDKit\lib\src\sl_core\utils\util.cu:482 code=700(cudaErrorIllegalAddress) “cudaCreateTextureObject(tex, &resDesc, &texDesc, NULL)”
in enum sl::ERROR_CODE __cdecl sl::Mat::updateCPUfromGPU(struct CUstream_st *) : Err [700]: an illegal memory access was encountered.
in enum sl::ERROR_CODE __cdecl sl::Mat::updateCPUfromGPU(struct CUstream_st *) : Err [700]: an illegal memory access was encountered.
CUDA error at C:\builds\sl\ZEDKit\lib\src\sl_zed\DetectionHandler.cpp:2283 code=700(cudaErrorIllegalAddress) “cudaStreamWaitEvent(cu_strm, ev_depth, 0)”
in enum sl::ERROR_CODE __cdecl sl::Mat::updateCPUfromGPU(struct CUstream_st *) : Err [700]: an illegal memory access was encountered.
CUDA error at C:\builds\sl\ZEDKit\lib\src\sl_zed\DetectionHandler.cpp:2315 code=700(cudaErrorIllegalAddress) “cudaStreamWaitEvent(cu_strm, ev_bt, 0)”
in enum sl::ERROR_CODE __cdecl sl::Mat::copyTo(class sl::Mat &,enum sl::COPY_TYPE,struct CUstream_st *) const : Err [700]: an illegal memory access was encountered.
in enum sl::ERROR_CODE __cdecl sl::Mat::copyTo(class sl::Mat &,enum sl::COPY_TYPE,struct CUstream_st *) const : Err [700]: an illegal memory access was encountered.
in enum sl::ERROR_CODE __cdecl sl::Mat::copyTo(class sl::Mat &,enum sl::COPY_TYPE,struct CUstream_st *) const : Err [700]: an illegal memory access was encountered.
CUDA error at C:\builds\sl\ZEDKit\lib\src\sl_zed\DetectionHandler.cpp:2157 code=700(cudaErrorIllegalAddress) “cudaEventRecord(ev_image, sdk_stream)”
CUDA error at C:\builds\sl\ZEDKit\lib\src\sl_zed\DetectionHandler.cpp:2245 code=700(cudaErrorIllegalAddress) “cudaStreamWaitEvent(cu_strm, ev_image, 0)”
in enum sl::ERROR_CODE __cdecl sl::Mat::copyTo(class sl::Mat &,enum sl::COPY_TYPE,struct CUstream_st *) const : Err [700]: an illegal memory access was encountered.
CUDA error at C:\builds\sl\ZEDKit\lib\src\sl_zed\DetectionHandler.cpp:2187 code=700(cudaErrorIllegalAddress) “cudaEventRecord(ev_depth, sdk_stream)”
[2025-11-23 18:47:37 UTC][ZED][WARNING] IExecutionContext::enqueueV3: Error Code 1: Cuda Runtime (In genericReformat::copyPackedRunKernel at C:_src\runtime\gpu\legacy\genericReformat.cuh:1501)[2025-11-23 18:47:37 UTC][ZED][WARNING] [Grab] CUDA error returned: an illegal memory access was encountered ( 700 )

Our Rig is a Blackwell GPU with i9:
NVIDIA Systeminformationen-Bericht erstellt am: 11/23/2025 19:31:32
Version der NVIDIA App: 11.0.5.420
Betriebssystem: Microsoft Windows 11 Pro, Version 10.0.26100
DirectX-Laufzeitversion: DirectX 12
Treiber: Game Ready Treiber - 581.80 - Tue Nov 4, 2025
CPU: Intel(R) Core™ i9-14900KF
RAM: 32.0 GB
Speicher: SSD - 1.8 TB

Grafikkarte
GPU: NVIDIA GeForce RTX 5080
Direct3D-Funktionsebene: 12_1
CUDA-Kerne: 10752
Grafik-Taktfrequenz: 2775 MHz
Resizable BAR: Ja
Speicher-Datenrate: 30.00 Gbps
Speicherschnittstelle: 256-Bit
Speicherbandbreite: 960.064 GB/s
Gesamter verfügbarer Grafikspeicher: 32581 MB
System-Videospeicher: Nicht anwendbar
Gemeinsam genutzter Systemspeicher: 16278 MB
Dedizierter Videospeicher: 16303 MB GDDR7
Video-BIOS-Version: 98.03.3b.00.a8
Geräte-ID: 10DE 2C02 1435196E
Teilenummer: G147 0045
IRQ: Nicht benutzt
Bus: PCI Express x16

Anzeige (1): Samsung LS27A80
Auflösung: 4k x 2k, 2560 x 1600 (empfohlen)
Bildwiederholfrequenz: 60 Hz
Desktop-Farbtiefe: Maximum (32 Bit)
HDCP: Unterstützt

i cant share our C++ public, but if possible we like to show you the code. Maybe we do something wrong…

Kindest Regards

Pascal

BenjaminV · November 24, 2025, 3:08pm

Hi,
You can share your code by sending us a mail at support@stereolabs.com so we can take a look a it .

pachermann · November 26, 2025, 9:34am

We have sent you the Email to Support with the Comitted Branch we worked on.

BenjaminV · November 26, 2025, 4:05pm

Thanks, I see it.

I’ll take a look and come back to you!

pachermann · December 5, 2025, 8:46am

Hi @BenjaminV any News on our code?

Can we use this same Fusion Program also on a older SDK with a non Blackwell GPU without changing it?

We need to start testing in the Museum and hope we can achieve a better result with it.

But on the long run we like to have a stable app or they will cancel our installation eventually

BenjaminV · December 5, 2025, 9:06am

Hi,

Sorry for not getting back to you sooner.

I reviewed your code and did not find any issues with it. I even ran it for multiple hours without any problem with similar hardware.
You are using the ZED SDK 5.1, is that correct?

pachermann · December 5, 2025, 9:41am

Yes on that PC with the crash, could it be the Crash of our Fusion is being triggered by the Unity Exe?
Or has it to be our rig then?

BenjaminV · December 5, 2025, 10:10am

The error seems to be related to a Cuda context inside the ZED SDK, I don’t think using Unity or not matters here. I think it’s on the “sender” side, not on the fusion side.

pachermann · December 5, 2025, 10:12am

Im sorry im stuck how can i check what SDK version my dev has installed?

BenjaminV · December 5, 2025, 10:27am

You can open ZED Explorer and look at the bottom right corner :

pachermann · December 5, 2025, 10:29am

thanks! It says v5.1.1

pachermann · December 5, 2025, 12:49pm

Im going to Wipe the whole machine next week and try setting up everything.

Can you maybe send us your specs?

We like to use the exact same Nvidia drivers for our blackwell GPU maybe we get lucky.

A few questions, if these could have an effect or be the cause:

only one harddrive, OS SDK & unity run from same harddrive SSD
wrong Nvidia driver versions? maybe the latest is not the best?
do we need to install cuda core sdk before stereolabs sdk?
anything that could lead to these errors when installing the machine with Windows 11?

BenjaminV · December 5, 2025, 1:12pm

Hi,

You need to install Cuda toolkit first, then the ZED SDK.

For ref, on my computer that uses a RTX 5090, I have CUDA 12.8 installed, and my driver version is 581.57.